Causal Thinking in Thermal Comfort
A well-known mantra in statistics says, “Correlation does not imply causation.” For instance, there might be a correlation between ice cream sales and shark attacks (both increase during the summer). However, it would be erroneous to infer that one causes the other. The aim of many scientific studies is to understand the causation between variables, but the statistical methods we commonly use are correlation-based. This may lead to the misinterpretation of correlation as causation, especially when we make assumptions about causal relationships based on plausible explanations. Causal thinking is a process to help disentangle correlations and causations.
Thermal comfort is a critical research topic in building science. Currently, 15% of global annual CO2 emissions come from maintaining indoor thermal comfort. While, on average, around 40% of occupants are dissatisfied with their indoor thermal environments. In thermal comfort studies, researchers commonly collect two types of data, one is occupants’ Thermal Sensation Vote (TSV), ranging from -3 (feeling hot) to 0 (feeling neutral) to +3(feeling hot), and another one is temperature measurement. After conducting a linear regression between two variables, researchers can calculate the neutral temperature by setting the TSV to zero.
In a recently published paper (open access), we applied causal thinking to the calculation of the neutral temperature; that is, the temperature at which a group of occupants would be experiencing a neutral sensation on average. There are two linear regression approaches used for calculating the neutral temperature:
-
Approach (a): the traditional approach, regresses TSV (y-axis) on indoor temperature (x-axis)
-
Approach (b): reverses the traditional approach and regresses indoor temperature on TSV.
In either approach, the neutral temperature can be calculated by setting the TSV to zero on the fitted regression line.
Some may think that the regression direction doesn’t matter and that the two approaches are interchangeable. However, causal thinking and the way the regression line is calculated reveal substantial and practical differences between them. Approach (a) represents the idea that the indoor thermal environment affects occupants’ thermal sensations. An example could be an occupant without opportunities to adjust the indoor temperature in a large open space office. In contrast, Approach (b) reflects situations where thermal sensations can trigger behavioral changes, not explicitly modeled here, which can then alter indoor thermal environments. An example could be an occupant opening a window or changing the thermostat in a private office.
Using the same data from one building as an illustrative example, we found that the two approaches lead to different ‘neutral temperature’ estimations. Approach (a), the most commonly used method, is how the original adaptive comfort models were developed. Approach (b) is less commonly used but aligns with the adaptive comfort theory that asserts building occupants are not merely passive recipients of their thermal environment; instead, it is more applicable in buildings where people have a high level of effective control of their thermal environment. We are currently exploring the implications of these different approaches in the larger ASHRAE Global Thermal Comfort Database II. The differences may have practical implications for the shape and temperature range of the adaptive thermal comfort models.
We use this simple example of a two-variable regression and data from a single illustrative building to highlight the importance of integrating causal thinking into correlation-based methods. Applying causal thinking to this topic led us to discuss the way neutral temperature is calculated, the underlying assumptions of regression methods, and the limitations of current field study data collection methods to allow us to better understand adaptive behavior.
Reference
- R. Sun, S. Schiavon, G. Brager, E. Arens, H. Zhang, T. Parkinson, and C. Zhang. Causal Thinking: Uncovering Hidden Assumptions and Interpretations of Statistical Analysis in Building Science. Building and Environment, (2024):111530. https://doi.org/10.1016/j.buildenv.2024.111530
Acknowledgment
The industry consortium members of the Center for the Built Environment (CBE) at the University of California, Berkeley, have supported this research.